Pith. sign in

REVIEW 3 major objections 3 minor 36 references

The paper proposes the value Causal Markov Condition (v-CMC), a normative principle asserting that in a causally sufficient DAG a variable's value is independent of its non-ancestors conditional on its causal children, and proves this yield

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 20:10 UTC pith:X7QCIZFU

load-bearing objection The v-CMC machinery is a real formal contribution, but the normative justification is close to circular—still deserves a serious referee. the 3 major comments →

arxiv 2607.16717 v1 pith:X7QCIZFU submitted 2026-07-18 stat.ML cs.AIcs.LG

A Causal Markov Condition for Value

Olav Benjamin Vassend This is my paper
classification stat.ML cs.AIcs.LG MSC 91B06
keywords value Causal Markov Conditionconditional value independencev-separationcausal DAGutility decompositionBellman recursionutility elicitationinfluence diagrams
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Value, the paper claims, should obey a causal independence principle just as probability does. It proposes the value Causal Markov Condition (v-CMC): in a causally sufficient DAG, once you fix the causal children of a variable, its conditional value contribution should not depend on any of its non-ancestors. The payoff is that a rational utility function then decomposes over the DAG into a sum of local child-conditional terms, a Bellman-style recursion on causal graphs, and a graphical criterion (v-separation) that is sound and complete for conditional value independence. If accepted, this gives a normative bridge between causal knowledge and preference, enabling modular transfer of utility information across contexts and structured elicitation procedures.

Core claim

The paper's central discovery is a duality between probability and value in causal graphs: probability flows downstream, value flows upstream. Formalizing this, it defines conditional value as subtraction, u(x|y)=u(x,y)-u(y), and conditional value independence accordingly, then states the local v-CMC: for any node Xi and any set N of non-ancestors, Xi is value-independent of N given its children Ch(Xi), whenever the graph is value-sufficient (all shared causal effects are included). The paper proves this local statement is equivalent to a global statement using v-separation (the exact dual of d-separation) and to a decomposition statement: over any child-closed subgraph, total utility is the

What carries the argument

The load-bearing objects are the subtractive conditional value u(x|y)=u(x,y)-u(y), which turns utilities into a semi-graphoid independence relation; the child-closed subgraph; and the translation key that swaps parents with children, non-descendants with non-ancestors, common causes with common effects, and p(A|B)=p(A,B)/p(B) with u(A|B)=u(A,B)-u(B). This key translates probabilistic CMC results into value results, and v-separation (defined as the value-dual of d-separation) provides the graphical criterion that is proven sound and complete for conditional value independence.

Load-bearing premise

The whole edifice is load-bearing on utilities that are defined on arbitrary subsets of the variables and are unique up to a common positive affine scale, so that the subtractive conditional value u(x|y)=u(x,y)-u(y) is meaningful; without that scale, the v-CMC cannot even be stated.

What would settle it

Construct a utility function on a small DAG (say three nodes with A→B←C) that satisfies the decomposition v-CMC but for which u(A | B, C) ≠ u(A | B); if such a function exists, the claimed equivalence between local and decomposition v-CMC fails, and the appendix's proof would be contradicted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Under the v-CMC, any rational utility function compatible with a value-sufficient DAG decomposes as u(V') = Σ u(Xi|Ch(Xi)) on every child-closed set, turning a joint elicitation problem into a sum of small local assessments.
  • v-separation gives a sound and complete graphical test for conditional value independence, so one can read utility independencies directly off the causal DAG.
  • Bellman recursion is the special case of the v-CMC decomposition on a linear chain when local conditional values do not depend on the children; the DAG version extends it to arbitrary causal graphs.
  • Updating utilities after an intervention or after adding a new common effect requires revising only the local terms whose child sets change, so the cost scales with the number of affected nodes, not the whole graph.
  • The decomposition licenses an elicitation algorithm whose number of required utility queries is at most n·2^{1+|Ch(W)|}, linear in the number of variables.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same probability-value duality may dualize further causal-inference concepts—transportability, confounding, counterfactuals—into value analogues, suggesting a broader program of causal value theory beyond the equivalences proven here.
  • Applied to inverse reinforcement learning, v-CMC could serve as a structural prior that selects among the many reward functions consistent with observed behavior; a natural test is whether recovered rewards on a known causal graph approximate the child-conditional decomposition.
  • The value-sufficiency requirement implies that practical elicitation protocols must explicitly include shared causal effects as variables; if common effects are omitted, the v-CMC will appear to be violated, which gives a diagnostic for model misspecification in preference elicitation.
  • The equivalence results depend on utilities being on a common interval or ratio scale across subsets; if only ordinal preferences are available, the v-CMC has no subtractive conditional value to constrain, so its normative force is tied to a cardinal utility interpretation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes a 'value Causal Markov Condition' (v-CMC): for a value-sufficient causal DAG and a utility function defined on subsets of variables, each node is conditionally value independent of its non-ancestors given its children (Definition 4). It argues this is a normative constraint on rational utility, introduces a probability--value duality (Section 5), and proves three equivalent formulations: local, global (v-separation), and decomposition (Definitions 10--12, Theorem 4). It also proves soundness and completeness of v-separation (Theorem 5), derives a Bellman-type recursion on DAGs (Theorem 6), and applies the decomposition to utility elicitation, modular transfer under interventions, and automatic influence-diagram construction (Section 7). The formal development is conditional on the v-CMC and on the assumption that utilities are defined on arbitrary subsets with a common positive affine scale.

Significance. The formal core is useful and mostly sound. The equivalence of local/global/decomposition versions is a natural dual of standard results, and the completeness proof using u = log P is a clean construction that does not rely on any fitted parameters. The Bellman-type decomposition, if correctly proved, gives a principled generalization of Bellman recursion to DAGs, and the elicitation and transfer algorithms are concrete and potentially practical. These conditional results are a genuine contribution to graphical utility models. The main weakness is that the paper's central normative claim -- that rational utility should obey the v-CMC -- is not independently established: the justification theorem derives the principle from premises that largely restate it. The paper would be publishable as a conditional mathematical framework, but not, in its current form, as an argument that rational preferences must satisfy the v-CMC.

major comments (3)
  1. [§4, Definitions 7--9 and Theorem 3] The normative justification is close to circular. Definition 7 (child mediation) requires W ⊥_u D' | Ch(W) for every non-child descendant set D', which is exactly the local v-CMC restricted to the descendant part of NA(W). In the minimal DAG W→Y→Z, child mediation is the whole v-CMC for W. The CDT property (Definition 9) supplies the remaining non-descendant part, but only for the interventional utility u_do(W); utility modularity (Definition 8) transfers that screening-off back to u. Thus Theorem 3 derives the v-CMC from premises that already contain the same screening-off intuition. The paper's caveats in Section 4 footnote 2 and Appendix A acknowledge scope limits, but they do not provide an independent argument that rational utility must screen off non-child descendants once children are fixed. Since the abstract claims v-CMC is a normative constraint, this is load-bearing. Please ei
  2. [§7.1, Theorem 6 and Appendix G, Eq. (37)] The proof of the key lemma (37) is under-specified. In the induction step, the proof uses the v-CMC to assert u({W} | Ch(W) ∪ S' ∪ Ch(S')) = u(W | Ch(W)) and u(S' | Ch(S') ∪ Ch(W)) = u(S' | Ch(S')). The conditioning set includes Ch(S'), which may contain nodes that are descendants of W, e.g., a common child of W and a node S' ∈ S. The local v-CMC (Definition 4) applies only to non-ancestors, and the stated reason 'W has no ancestors in S' does not rule out such descendants. The step can be repaired by invoking the global v-CMC and showing that S' ∪ Ch(S') is v-separated from W by Ch(W), but as written the proof is incomplete. This matters because Theorem 6 is the paper's advertised Bellman-type generalization.
  3. [§2, Definitions 1--3] The framework requires utilities to be defined on arbitrary subsets of variables and to be unique up to a common positive affine scale across all subsets, and it imposes conditional consistency (Definition 3) as an additional axiom. These are strong assumptions. The paper notes that it is framework-neutral and cites existing frameworks, but it does not establish that rational preferences in those frameworks always admit such a common-scale representation on all subsets of a causal variable set. If utilities are only ordinally comparable, or are defined only on acts, the v-CMC and all equivalences in Theorems 4--6 have no domain. The paper should state this limitation prominently and specify which decision-theoretic settings satisfy the required scale condition.
minor comments (3)
  1. [Appendix D] The symbol ND is used both for N ∩ (D \ Ch(W)) and for the set of all non-descendants of W in Gdo(W). This makes the proof of Theorem 3 hard to follow and should be disambiguated.
  2. [Appendix D, after Eq. (17)] The claim that 'the graphical relations between W, NND, and D are not affected by the intervention do(W)' is imprecise. The needed argument is that utility modularity (Definition 8) applies to the specific subsets that enter the conditional-value expression; please spell this out.
  3. [Appendix F] In the completeness proof, the constructed value function u = log P is a formal witness only; it is not claimed to be a normatively justified utility. Stating this explicitly would avoid unnecessary objections.

Circularity Check

2 steps flagged

Normative derivation of the v-CMC is partly circular: Theorem 3's child-mediation premise is the v-CMC restricted to descendants, and the CDT property is the same causal-effects screening-off intuition for interventions.

specific steps
  1. self definitional [Section 4, Definition 7 (Child mediation) and Definition 4 (local v-CMC), used in Theorem 3]
    "Definition 7: u is child-mediated with respect to G iff, for every node W and for every set D′ of causal descendants of W that are not children of W, W⊥⊥u D′ |Ch(W). Definition 4: for any set of non-ancestors N⊂NA(Xi): Xi ⊥⊥u N|Ch(Xi)."

    Every non-child descendant of W is a non-ancestor of W, so Definition 7 is exactly the local v-CMC of Definition 4 restricted to descendant-only N. In a minimal chain W→Y→Z, NA(W)={Z}, so child mediation is the whole v-CMC for W. Theorem 3 then proves the v-CMC by using, as a premise, the v-CMC itself on the descendant part of the conditioning problem (proof step (15): 'Using child mediation...'). The normative content of the conclusion is thus assumed in the premise rather than derived from weaker or independent assumptions.

  2. other [Section 4, Definition 9 (CDT property) and Theorem 3]
    "Definition 9: u has the CDT property with respect to G iff for any ND′ ⊂ND: W⊥⊥_udo(W) ND′ |D. ... it simply requires that the utility of an intervention be assessed on the basis of its causal effects alone, with other variables relevant only insofar as they causally influence those effects."

    The CDT property is an interventional screening-off condition: under do(W), non-descendants are value-irrelevant once all descendants are fixed. By utility modularity (Definition 8), this transfers to u as W⊥⊥_u ND′ |D, which is the same 'value depends only on causal effects' screening-off idea that the v-CMC is introduced to formalize, with the conditioning set expanded from Ch(W) to all descendants. Theorem 3 therefore relies on the very causal-effects-screening-off intuition it claims to justify, making the normative defence of the v-CMC substantially circular even though the subsequent formal equivalences are non-circular.

full rationale

The paper's formal contributions are largely self-contained: Theorems 1, 2, 4, 5, and 6 are proved from stated definitions, and the soundness/completeness result for v-separation is obtained by an explicit log-probability construction that imports standard d-separation completeness without assuming the v-CMC. There is no data fitting, and no 'prediction' is declared from fitted parameters. Self-citation is not load-bearing. However, the advertised normative justification of the v-CMC (Section 4, Theorem 3) is partly circular. Definition 7 (child mediation) is the local v-CMC of Definition 4 restricted to descendant non-children, and Definition 9 (CDT property) is the same causal-effects screening-off idea stated in interventional form. Thus the theorem that is presented as the main 'defense' of the v-CMC packages the principle into its premises rather than deriving it from independent normative foundations. This warrants a moderate circularity score. The equivalence and decomposition results remain independent mathematical content, so the score is not higher.

Axiom & Free-Parameter Ledger

0 free parameters · 7 axioms · 0 invented entities

The central claim rests on a stack of domain assumptions: a common interval utility scale, the subtractive definition of conditional value, conditional consistency, and value sufficiency. The normative derivation additionally assumes child mediation, utility modularity, and the CDT property. The formal theorems then rely on standard graphoid results.

axioms (7)
  • domain assumption Utility functions are unique up to positive affine transformations on a common scale across all variable subsets (Definition 1).
    Without a common interval scale, the subtractive conditional value u(x|y)=u(x,y)-u(y) and the value independence relation are not well-defined; the paper invokes this in Definition 1.
  • domain assumption Conditional value is defined by subtraction: u(x|y)=u(x,y)-u(y) (Eq. 2).
    The v-CMC is stated in terms of this Brafman-Engel definition; Theorem 2 gives a uniqueness argument but under assumptions that are themselves modeling choices.
  • domain assumption Conditional consistency: (A ⊥⊥_u B|C) implies (A ⊥⊥_u B'|C) for B'⊆B (Definition 3).
    Imposed to make conditional value independence a semi-graphoid; needed for the local/global equivalence proof.
  • domain assumption Value sufficiency: all shared causal effects of any two variables are included in the DAG (Definition 5).
    The compatibility criterion and the v-CMC apply only to value-sufficient DAGs; omission of a shared effect breaks the v-CMC (Section 4).
  • domain assumption The causal DAG correctly represents the causal structure of the variables.
    All theorems presuppose a correct causal graph; the v-CMC is normative relative to that graph.
  • ad hoc to paper Child mediation, utility modularity, and the CDT property are acceptable normative premises (Definitions 7-9).
    Theorem 3 derives v-CMC from these; they are close to the conclusion and are not independently established.
  • standard math Standard graphoid results: local/global CMC equivalence for semi-graphoids (Lauritzen Thm 2.44); d-separation completeness (Geiger et al.); faithfulness is generic (Meek).
    Used in the proofs of Theorems 4 and 5.

pith-pipeline@v1.3.0-alltime-deepseek · 21671 in / 32325 out tokens · 285554 ms · 2026-08-01T20:10:57.054146+00:00 · methodology

0 comments
read the original abstract

This paper proposes a causal independence principle for value -- the value Causal Markov Condition (v-CMC) -- and develops the conceptual and mathematical foundations of a "causal value theory" linking causality and utility. After motivating a local formulation of the v-CMC, we introduce a probability-value duality that translates standard causal-inference results into the value setting. In particular, we formulate local, global, and decomposition versions of the v-CMC and prove their equivalence. We also define v-separation and show that it is sound and complete for conditional value independence. Furthermore, we derive a Bellman-type recursion as a special case of the v-CMC, thereby generalizing standard Bellman recursion from linear chains to causal DAGs. Finally, we show how the v-CMC supports modular transfer and updating of utility information across causal contexts and develop algorithms for causally structured utility elicitation and canonical influence-diagram construction.

Figures

Figures reproduced from arXiv: 2607.16717 by Olav Benjamin Vassend.

Figure 1
Figure 1. Figure 1: Antibiotic example: A affects S, E, which affect outcomes T, W; D, I, R are exogenous. Consider A in [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Antibiotic example with value nodes attached to [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Post-intervention graph GA: A affects S, E, which affect outcomes T, W; D is disconnected and I, R are exogenous. u A(A | S, E, D, I, R, T, W) = u A(A | S, E). (49) Hence, although the original graph contains eight variables, the local conditional-value assessment of administering the antibiotic requires only that we condition on the pair (S, E). This illustrates how v-separation can simplify local conditi… view at source ↗
Figure 4
Figure 4. Figure 4: Post-intervention graph GAE: A affects S, while E is fixed at 0 by intervention. The intervention on E affects the child sets of both R and A. However, R is not part of the reduced child-closed set {A, S, E, T, W} used above, and its local term is therefore not needed for the present analysis. Since we are concerned with the downstream utility of A, and E is fixed at 0 and is no longer part of the minimal … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references

  1. [1]

    Adams, S., Cody, T., and Beling, P. A. (2022). A survey of inverse reinforcement learning. Artificial Intelligence Review , 55:4307--4346

    2022

  2. [2]

    Anscombe, F. J. and Aumann, R. J. (1963). A definition of subjective probability. Annals of Mathematical Statistics , 34:199--205

    1963

  3. [3]

    and Doshi, P

    Arora, S. and Doshi, P. (2021). A survey of inverse reinforcement learning: Challenges, methods and progress. Artificial Intelligence , 297:103500

    2021

  4. [4]

    and Grove, A

    Bacchus, F. and Grove, A. J. (1995). Graphical models for preference and utility. In Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence (UAI-95) , pages 3--10, San Francisco. Morgan Kaufmann

    1995

  5. [5]

    Bellman, R. (1957). Dynamic Programming . Princeton University Press, Princeton, NJ

    1957

  6. [6]

    Bolker, E. D. (1967). A simultaneous axiomatization of utility and subjective probability. Philosophy of Science , 34(4):333--340

    1967

  7. [7]

    Boutilier, C., Bacchus, F., and Brafman, R. I. (2001). UCP -networks: A directed graphical representation of conditional utilities. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence , pages 56--64, San Francisco. Morgan Kaufmann

    2001

  8. [8]

    I., Domshlak, C., Hoos, H

    Boutilier, C., Brafman, R. I., Domshlak, C., Hoos, H. H., and Poole, D. (2004). CP -nets: A tool for representing and reasoning with conditional ceteris paribus preference statements. Journal of Artificial Intelligence Research , 21:135--191

    2004

  9. [9]

    Bradley, R. (2017). Decision Theory with a Human Face . Cambridge University Press

    2017

  10. [10]

    Brafman, R. I. and Engel, Y. (2009). Conditional utility independence: A new notion with applications in multiattribute utility theory. In Proceedings of the Twenty-First International Joint Conference on Artificial Intelligence (IJCAI 2009) , pages 389--394

    2009

  11. [11]

    Brafman, R. I. and Engel, Y. (2010). Decomposed utility functions and graphical models for reasoning about preferences. In Proceedings of the 24th AAAI Conference on Artificial Intelligence , 267--272

    2010

  12. [12]

    and Wellman, M

    Engel, Y. and Wellman, M. P. (2008). CUI networks: A graphical representation for conditional utility independence. Journal of Artificial Intelligence Research , 31:83--112

    2008

  13. [13]

    D., Ortega, P

    Everitt, T., Carey, R., Langlois, E. D., Ortega, P. A., and Legg, S. (2021). Agent incentives: A causal perspective. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35, pages 11487--11495

    2021

  14. [14]

    and Hutter, M

    Everitt, T. and Hutter, M. (2021). Causal influence diagrams for safe and fair AI . Artificial Intelligence , 296:103479

    2021

  15. [15]

    Fishburn, P. C. (1974). Seven independence concepts and continuous multiattribute utility functions. Journal of Mathematical Psychology , 11(3):337--359

    1974

  16. [16]

    Geiger, D., Verma, T., and Pearl, J. (1990a). d-separation: From theorems to algorithms. In Proceedings of the Fifth Conference on Uncertainty in Artificial Intelligence , pages 139--148. Elsevier

  17. [17]

    S., and Pearl, J

    Geiger, D., Verma, T. S., and Pearl, J. (1990b). Identifying independence in Bayesian networks. Networks , 20(5):507--534

  18. [18]

    and Harper, W

    Gibbard, A. and Harper, W. L. (1978). Counterfactuals and two kinds of expected utility. In Hooker, C., Leach, J. J., and McClennen, E. F., editors, Foundations and Applications of Decision Theory , pages 125--162. Reidel, Dordrecht

    1978

  19. [19]

    Howard, R. A. and Matheson, J. E. (2005). Influence diagrams. Decision Analysis , 2(3):127--143

    2005

  20. [20]

    Jeffrey, R. C. (1965). The Logic of Decision . McGraw-Hill

    1965

  21. [21]

    Jeffrey, R. C. (1983). The Logic of Decision . University of Chicago Press, 2nd edition

    1983

  22. [22]

    Joyce, J. M. (1999). The Foundations of Causal Decision Theory . Cambridge University Press, Cambridge

    1999

  23. [23]

    Keeney, R. L. and Raiffa, H. (1993). Decisions with Multiple Objectives: Preferences and Value Tradeoffs . Cambridge University Press, Cambridge, 2nd edition

    1993

  24. [24]

    Lauritzen, S. L. (2019). Lectures on graphical models. 3rd electronic edition (lecture notes)

    2019

  25. [25]

    and Smith, J

    Leonelli, M. and Smith, J. Q. (2017). Directed expected utility networks. Decision Analysis , 14(2):108--125

    2017

  26. [26]

    Lewis, D. (1981). Causal decision theory. Australasian Journal of Philosophy , 59(1):5--30

    1981

  27. [27]

    Meek, C. (1995). Strong completeness and faithfulness in Bayesian networks. In Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence , pages 411--418

    1995

  28. [28]

    Mura, P. L. and Shoham, Y. (1999). Expected utility networks. In Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence (UAI-99) , pages 366--373, San Francisco. Morgan Kaufmann

    1999

  29. [29]

    Pearl, J. (2009). Causality: Models, Reasoning, and Inference . Cambridge University Press, Cambridge, 2nd edition

    2009

  30. [30]

    and Paz, A

    Pearl, J. and Paz, A. (1987). Graphoids: A graph-based logic for reasoning about relevance relations. In du Boulay, B., Hogg, D., and Steels, L., editors, Advances in Artificial Intelligence-II , pages 357--363. North-Holland

    1987

  31. [31]

    Savage, L. J. (1954). The Foundations of Statistics . Wiley

    1954

  32. [32]

    Shachter, R. D. (1986). Evaluating influence diagrams. Operations Research , 34(6):871--882

    1986

  33. [33]

    N., and Scheines, R

    Spirtes, P., Glymour, C. N., and Scheines, R. (2000). Causation, Prediction, and Search . MIT Press, Cambridge, MA, 2nd edition

    2000

  34. [34]

    Stern, R. (2017). Interventionist decision theory. Synthese , 194:4133--4153

    2017

  35. [35]

    and Pearl, J

    Verma, T. and Pearl, J. (1988). Causal networks: Semantics of conditional independence. In Proceedings of the Fourth Conference on Uncertainty in Artificial Intelligence , pages 69--78

    1988

  36. [36]

    and Morgenstern, O

    von Neumann, J. and Morgenstern, O. (1944). Theory of Games and Economic Behavior . Princeton University Press

    1944